dhruvkys/hi-asr-1k

Hindi ASR 1k

A small, ready-to-use Hindi automatic speech recognition (ASR) dataset: ~2,000 short read-speech utterances with ground-truth Devanagari transcriptions, derived from Mozilla Common Voice (Hindi). It is sized for quick fine-tuning experiments and for benchmarking/word-error-rate (WER) evaluation of models such as Whisper and other multilingual/Indic ASR systems — small enough to iterate on a single GPU or even CPU, while still being a real, human-spoken evaluation set.

• Language: Hindi (hi), Devanagari script
• Task: Automatic Speech Recognition (speech → text)
• Audio: MP3, mono
• Splits: train (1,000 transcribed clips) / test (1,000 transcribed clips)
• Source: Mozilla Common Voice Hindi (crowd-sourced read speech)
• License: CC0-1.0 (public domain), inherited from Common Voice

Dataset structure

Each split ships audio under clips/ (original Common Voice filenames) with a metadata file mapping every clip to its transcript. A parallel audio/ folder holds the same kind of clips under sequential cv_NNNNNN.mp3 identifiers.

hi-asr-1k/
├── train/
│   ├── clips/           # 1,000 common_voice_hi_*.mp3
│   ├── audio/           #   800 cv_*.mp3  (sequential IDs)
│   └── metadata.tsv     # file_name, text  (transcripts for the clips/)
└── test/
    ├── clips/           # 1,000 common_voice_hi_*.mp3
    ├── audio/           #   200 cv_*.mp3  (sequential IDs)
    ├── metadata.csv     # file_name, text  (transcripts for the clips/)
    └── metadata.tsv     # path, text       (transcripts for the audio/ files)

Data fields

• file_name / path (string): relative path to the audio clip.
• text (string): the ground-truth Hindi transcription in Devanagari.
• audio (Audio): the decoded waveform + sampling rate, when loaded via the 🤗 datasets AudioFolder builder.

Data instance

{
  "file_name": "common_voice_hi_24663092.mp3",
  "text": "उन्हे अपने अकलमंद बेटे पर नाज़ है।"
}

Splits

Split	Transcribed clips (clips/)	Extra audio (audio/)
train	1,000	800
test	1,000	200

Note: train/metadata.tsv lists 9,000 Common Voice transcripts as a reference superset; only the 1,000 clips actually shipped in train/clips/ have audio.

Usage

Because the audio lives in per-split subfolders, load it with the audiofolder builder and point it at the split and its metadata:

from datasets import load_dataset, Audio

# Load the transcribed clips for one split
ds = load_dataset(
    "audiofolder",
    data_dir="test",              # or "train"
    split="train",                # audiofolder names the single split "train"
)
ds = ds.cast_column("audio", Audio(sampling_rate=16_000))

print(ds[0]["text"])
print(ds[0]["audio"]["array"].shape, ds[0]["audio"]["sampling_rate"])

Or read the transcripts directly with pandas:

import pandas as pd

df = pd.read_csv(
    "hf://datasets/dhruvkys/hi-asr-1k/test/metadata.csv",
    sep="\t",                     # metadata is tab-separated
)
print(df.head())

Evaluating a Whisper model (WER)

import evaluate
from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
               model="openai/whisper-small", generate_kwargs={"language": "hi"})
wer = evaluate.load("wer")

preds = [asr(x["audio"])["text"] for x in ds]
refs  = [x["text"] for x in ds]
print("WER:", wer.compute(predictions=preds, references=refs))

Source and collection

Audio and transcriptions originate from the Mozilla Common Voice Hindi corpus — sentences read aloud and contributed by volunteers, then validated by community review. This dataset is a curated ~1k-per-split subset repackaged for convenient fine-tuning and evaluation; no new recordings were made.

Preprocessing recommendations

• Common Voice clips are typically 8–48 kHz MP3; resample to 16 kHz mono for Whisper and most ASR models (shown above via Audio(sampling_rate=16000)).
• Transcriptions preserve original casing, punctuation, and occasional Latin-script tokens (e.g. brand/product names in headlines). For WER, consider normalizing punctuation and Latin numerals/tokens depending on your evaluation protocol.

Licensing

Released under CC0-1.0 (public domain dedication), consistent with the Mozilla Common Voice license. You may use, modify, and redistribute the data, including commercially, without restriction. Attribution to Mozilla Common Voice is appreciated but not required.

Citation

If you use this dataset, please cite Mozilla Common Voice:

@inproceedings{commonvoice,
  author    = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and
               Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and
               Tyers, F. M. and Weber, G.},
  title     = {Common Voice: A Massively-Multilingual Speech Corpus},
  booktitle = {Proceedings of the 12th Conference on Language Resources and
               Evaluation (LREC 2020)},
  pages     = {4211--4215},
  year      = {2020}
}

Known limitations

• Small scale. ~1k clips per split — good for prototyping and benchmarking, not for training a production ASR model from scratch.
• Domain skew. Many sentences are news headlines, so vocabulary leans toward named entities, politics, and current-affairs terms.
• Two naming schemes. clips/ uses original Common Voice filenames; audio/ uses sequential cv_* IDs. The train/audio/ clips do not ship a dedicated transcript file (only test/audio/ has test/metadata.tsv).
• Metadata format. The metadata files are tab-separated, even the one named metadata.csv. Pass sep="\t" when reading them (see usage above).